Megatest

		      " -server " (or target-host "-") (if (equal? (configf:lookup *configdat* "server" "daemonize") "yes")
							   " -daemonize "
							   "")
		      ;; " -log " logfile
		      " -m testsuite:" testsuite
		      " " profile-mode
		      )) ;; (conc " >> " logfile " 2>&1 &")))))
	 (log-rotate  (make-thread common:rotate-logs  "server run, rotate logs thread")) ;; why are we rotating logs here? This is a sensitive location with a lot going on!?
         (load-limit  (configf:lookup-number *configdat* "jobtools" "max-server-start-load" default: 3.0)))

    ;; we want the remote server to start in *toppath* so push there
    (push-directory areapath)
    (debug:print 0 *default-log-port* "INFO: Trying to start server (" cmdln ") ...")
    (thread-start! log-rotate)
    
    ;; host.domain.tld match host?
    (if (and target-host 
	     ;; look at target host, is it host.domain.tld or ip address and does it 
	     ;; match current ip or hostname
	     (not (string-match (conc "("curr-host "|" curr-host"\\..*)") target-host))
	     (not (equal? curr-ip target-host)))
	(begin
	  (debug:print-info 0 *default-log-port* "Starting server on " target-host ", logfile is " logfile)
	  (setenv "TARGETHOST" target-host)))
      
    (setenv "TARGETHOST_LOGF" logfile)
    (thread-sleep! (/ (random 5000) 1000)) ;; add about a random (up to 5 seconds) initial delay. It seems pretty common that many running tests request a server at the same time
    ;; (common:wait-for-normalized-load load-limit " delaying server start due to load" target-host) ;; do not try starting servers on an already overloaded machine, just wait forever
    #;(common:wait-for-homehost-load load-limit (conc " delaying server start due to load on homehost. limit is " load-limit))
    (system (conc "nbfake " cmdln))
    (unsetenv "TARGETHOST_LOGF")
    (if (get-environment-variable "TARGETHOST")(unsetenv "TARGETHOST"))
    (thread-join! log-rotate)
    (pop-directory)))

;; given a path to a server log return: host port startseconds
;;
(define (server:logf-get-start-info logf)
  (let ((rx (regexp "^SERVER STARTED: (\\S+):(\\d+) AT ([\\d\\.]+)"))) ;; SERVER STARTED: host:port AT timesecs
    (handle-exceptions
	exn
      (begin
	(print "failed to get server info from " logf ", exn=" exn)
	(list #f #f #f)) ;; no idea what went wrong, call it a bad server
      (with-input-from-file
	  logf
	(lambda ()
	  (let loop ((inl  (read-line))
		     (lnum 0))
	    (if (not (eof-object? inl))
		(let ((mlst (string-match rx inl)))

    ;; otherwise attempt to create the logs dir and then
    ;; continue
    (if (if (directory-exists? (conc areapath "/logs"))
	    '()
	    (if (file-write-access? areapath)
		(begin
		  (condition-case
		   (create-directory (conc areapath "/logs") #t)
		   (exn (i/o file)(debug:print 0 *default-log-port* "ERROR: Cannot create directory at " (conc areapath "/logs")))
		   (exn ()(debug:print 0 *default-log-port* "ERROR: Unknown error attemtping to get server list. exn=" exn)))
		  (directory-exists? (conc areapath "/logs")))
		'()))
	(let* ((server-logs   (glob (conc areapath "/logs/server-*.log")))
	       (num-serv-logs (length server-logs)))
	  (if (null? server-logs)
	      '()
	      (let loop ((hed  (car server-logs))
			 (tal  (cdr server-logs))
			 (res '()))
		(let* ((mod-time  (handle-exceptions
				   exn
				   (begin
				     (print "failed to get modification time on " hed ", exn=" exn)
				     (current-seconds)) ;; 0
				   (file-modification-time hed))) ;; default to *very* old so log gets ignored if deleted
		       (down-time (- (current-seconds) mod-time))
		       (serv-dat  (if (or (< num-serv-logs 10)
				  	  (< down-time 900)) ;; day-seconds))
				      (server:logf-get-start-info hed)
				      '())) ;; don't waste time processing server files not touched in the 15 minutes if there are more than ten servers to look at
		       (serv-rec (cons mod-time serv-dat))
		       (fmatch   (string-match fname-rx hed))
		       (pid      (if fmatch (string->number (list-ref fmatch 2)) #f))
		       (new-res  (if (null? serv-dat)
				     res
				     (cons (append serv-rec (list pid)) res))))
		  (if (null? tal)
		      (if (and limit
			       (> (length new-res) limit))
			  new-res ;; (take new-res limit)  <= need intelligent sorting before this will work
			  new-res)
		      (loop (car tal)(cdr tal) new-res)))))))))

(define (server:get-num-alive srvlst)
  (let ((num-alive 0))
    (for-each
     (lambda (server)
       (match-let (((mod-time host port start-time pid)
		    server))

				   (> (length rec) 2))
			      (let ((start-time (list-ref rec 3))
				    (mod-time   (list-ref rec 0)))
				;; (print "start-time: " start-time " mod-time: " mod-time)
				(and start-time mod-time
				     (> (- now start-time) 0)    ;; been running at least 0 seconds
				     (< (- now mod-time)   16)   ;; still alive - file touched in last 16 seconds
				     (or (not (configf:lookup *configdat* "server" "runtime")) ;; skip if not set
					 (< (- now start-time)       
					    (+ (- (string->number (configf:lookup *configdat* "server" "runtime"))
					       180)
					    (random 360)))) ;; under one hour running time +/- 180
				     ))
			      #f))
			srvlst)
		(lambda (a b)
		  (< (list-ref a 3)
		     (list-ref b 3))))))
    (if (> (length slst) nums)

(define (server:get-client-signature) ;; BB> why is this proc named "get-"?  it returns nothing -- set! has not return value.
  (if *my-client-signature* *my-client-signature*
      (let ((sig (server:mk-signature)))
        (set! *my-client-signature* sig)
        *my-client-signature*)))

;; wait for server=start-last to be three seconds old
;;
(define (server:wait-for-server-start-last-flag areapath)
  (let* ((start-flag (conc areapath "/logs/server-start-last"))
	 ;;; THIS INTERACTS WITH [server] timeout. Suggest using 0.1 or above for timeout (6 seconds)
	 (reftime    (configf:lookup-number *configdat* "server" "idletime" default: 4))
	 (server-key (conc (get-host-name) "-" (current-process-id))))
    (if (file-exists? start-flag)
	(let* ((fmodtime (file-modification-time start-flag))
	       (delta    (- (current-seconds) fmodtime))
	       (all-go   (> delta reftime)))
	  (if (and all-go
		   (begin
		     (with-output-to-file start-flag
		       (lambda ()
			 (print server-key)))
		     (thread-sleep! 0.25)
		     (let ((res (with-input-from-file start-flag
				  (lambda ()
				    (read-line)))))
		       (equal? server-key res))))
	      #t ;; (system (conc "touch " start-flag)) ;; lazy but safe
	      (begin
		(debug:print-info 0 *default-log-port* "Gating server start, last start: "
				  fmodtime ", delta: " delta ", reftime: " reftime ", all-go=" all-go)
		(thread-sleep! reftime)
		(server:wait-for-server-start-last-flag areapath)))))))

;; kind start up of servers, wait 40 seconds before allowing another server for a given
;; run-id to be launched
;;
(define (server:kind-run areapath)
  ;; look for $MT_RUN_AREA_HOME/logs/server-start-last
  ;; and wait for it to be at least 3 seconds old
  (server:wait-for-server-start-last-flag areapath)
  (if (not (server:check-if-running areapath)) ;; why try if there is already a server running?
      (let* ((last-run-dat (hash-table-ref/default *server-kind-run* areapath '(0 0))) ;; callnum, whenrun
	     (call-num     (car last-run-dat))
	     (when-run     (cadr last-run-dat))
	     (run-delay    (+ (case call-num
				((0)    0)
				((1)   20)
				((2)  300)
				(else 600))
			      (random 5)))   ;; add a small random number just in case a lot of jobs hit the work hosts simultaneously
	     (lock-file    (conc areapath "/logs/server-start.lock")))
	(if	(> (- (current-seconds) when-run) run-delay)
		(let* ((start-flag (conc areapath "/logs/server-start-last")))
		  (common:simple-file-lock-and-wait lock-file expire-time: 15)
		  (system (conc "touch " start-flag)) ;; lazy but safe
		  (server:run areapath)
		  (thread-sleep! 2) ;; don't release the lock for at least a few seconds
		  (common:simple-file-release-lock lock-file)))
	(hash-table-set! *server-kind-run* areapath (list (+ call-num 1)(current-seconds))))))

;; this one seems to be the general entry point
;;
(define (server:start-and-wait areapath #!key (timeout 60))
  (let ((give-up-time (+ (current-seconds) timeout)))
    (let loop ((server-url (server:check-if-running areapath))
	       (try-num    0))
      (if (or server-url
	      (> (current-seconds) give-up-time)) ;; server-url will be #f if no server available.
	  server-url

(define (server:writable-watchdog-deltasync dbstruct)
  (thread-sleep! 0.05) ;; delay for startup
  (let ((legacy-sync  (common:run-sync?))
        (sync-stale-seconds (configf:lookup-number *configdat* "server" "sync-stale-seconds" default: 300))
	(debug-mode   (debug:debug-mode 1))
	(last-time    (current-seconds))
	(no-sync-db   (db:open-no-sync-db))
	(stmt-cache   (dbr:dbstruct-stmt-cache dbstruct))
        (sync-duration 0) ;; run time of the sync in milliseconds
        ;;(this-wd-num  (begin (mutex-lock! *wdnum*mutex) (let ((x *wdnum*)) (set! *wdnum* (add1 *wdnum*)) (mutex-unlock! *wdnum*mutex) x)))
        )
    (set! *no-sync-db* no-sync-db) ;; make the no sync db available to api calls
    (debug:print-info 2 *default-log-port* "Periodic sync thread started.")
    (debug:print-info 3 *default-log-port* "watchdog starting. legacy-sync is " legacy-sync" pid="(current-process-id)  );;  " this-wd-num="this-wd-num)
    (if (and legacy-sync (not *time-to-exit*))

		  (if (and (not *time-to-exit*)
			   (< count 6)) ;; was 11, changing to 4. 
		      (begin
			(thread-sleep! 1)
			(delay-loop (+ count 1))))
		  (if (not *time-to-exit*) (loop))))
	    ;; time to exit, close the no-sync db here
	    (db:no-sync-close-db no-sync-db stmt-cache)
	    (if (common:low-noise-print 30)
		(debug:print-info 0 *default-log-port* "Exiting watchdog timer, *time-to-exit* = " *time-to-exit*" pid="(current-process-id) ))))))) ;;" this-wd-num="this-wd-num)))))))